Conditions Decide Whether AI Compounds or Costs

This is the last article in the series, and it is the one the others were building toward. Everything before it described a failure mode. This one describes the variable underneath all of them.

Two organizations

The first organization deploys AI across three functions. Usage climbs steadily. Engineers generate code faster than the review queue can absorb it. Analysts produce more analysis in a quarter than the previous team produced in a year. Marketing ships four times the volume of content. Every dashboard the program reports on is moving in the right direction.

Nothing a customer touches has changed. Cycle times are the same. Defect rates are the same. The analysis that gets produced does not reach anyone with the authority to act on it, so it accumulates. The code that gets generated does not get reviewed at the rate it arrives, so it ships with a lien on it. The content that gets published does not convert any better than the content that came before, because the constraint was never volume. Costs are up by the amount of the contract plus the amount of the rework.

The second organization deploys the same tools into the same three functions. Usage climbs at a similar rate. The difference does not show up in the consumption numbers at all, which is the first thing worth noticing about it.

It shows up eleven months later, in things that are visible from outside. A category of customer request that used to take four days is answered the same afternoon. A quarterly analysis cycle that occupied a team for three weeks now occupies two people for two days, and the reclaimed time went somewhere specific rather than dissolving into more meetings. A class of work the organization used to decline as uneconomic is now work it takes. Someone can name three things the organization stopped doing entirely.

The tools did none of that on their own. The tools produced output in both places. Only one of the two could convert it.

What follows is what actually differed, walked one condition at a time.

Attention: what the organization notices, it measures

Attention is the first condition because it determines what the organization is even capable of seeing, and an organization cannot manage what it has not noticed.

The first organization measured consumption. Not out of laziness, and not because anyone believed tokens were the point. It measured consumption because consumption was available on day one and outcomes were not, and because the board asked for evidence of return on a schedule that arrived long before any outcome could mature. Faced with that, most leadership teams reach for the number that exists.

The consequence is not that the number was wrong. The number was accurate. The consequence is that the organization spent a year looking at inputs, and everything that went wrong during that year went wrong in a place the organization was not looking. The review backlog, the decision queue, the accumulating maintenance liability, none of that appears on a usage chart. It was all happening. It was simply outside the field of view.

The second organization made a harder choice early. It accepted a reporting gap. For two quarters it had nothing impressive to show, because it had committed to measuring things that take two quarters to move. That decision was uncomfortable and it was the whole game. Once the organization was looking at outcomes, it could see which parts of the deployment were working and which were producing expensive motion, and it could move resources accordingly. The first organization could not see that, so it could not move anything.

Attention is not a reporting preference. It is the condition that determines whether the other three problems are visible at all.

Alignment: when productive means five different things

Ask five function heads what more productive with AI would look like in their area. In most organizations you will get five answers, and the answers will not be compatible with each other.

Engineering means throughput of shipped work. Legal means fewer unreviewed exposures, which is a brake on throughput by design. Finance means cost per unit of output. Sales means speed of response. Operations means predictability, which is often the opposite of speed. None of those definitions is wrong. Each one is a correct reading of what that function is accountable for.

The problem arrives when a general purpose capability is dropped on top of five definitions that were never reconciled. Each function optimizes locally, and each one succeeds locally. Engineering does ship more. Legal does review more thoroughly. Sales does respond faster. And the organization as a whole moves no better than it did before, because the local gains are pulling against each other and the friction absorbs the difference.

This is the oldest finding in the complementary assets literature and it long predates AI. Erik Brynjolfsson and Lorin Hitt argued in 2000 that the returns to information technology come not from the technology but from the organizational changes that accompany it: redesigned processes, redistributed decision rights, changed work practices. The computers were never the differentiator. What the firm did around the computers was.

The second organization did something unglamorous before it deployed anything. It got one definition of what better meant for the work in question, wrote it down, and made the functions argue about it in advance rather than discover the disagreement eleven months later in a variance report. That conversation took weeks and produced no output anyone could demo. It is the reason the local gains added up.

Authority: options are not decisions

AI is very good at generating options. It expands the space of things that could be done, and it does so faster than any prior tool.

An option only has value if something turns it into a decision. In the first organization, the generation rate went up by an order of magnitude and the decision rate did not move, because decision rights had not changed. The same approval chain, the same committee cadence, the same three people whose sign-off everything routes through. What had been a manageable queue became a backlog, and a backlog of unmade decisions is not neutral. It ages. Options expire, context decays, and the people who generated the work learn that generating it changes nothing.

That last effect is the expensive one. When capable people watch their output pile up unacted on, they stop producing the kind of output that requires judgment and start producing the kind that satisfies the dashboard. The organization gets exactly what it structurally rewarded.

The second organization moved decision rights down before it scaled the tooling. Not everything, and not dramatically. It identified the specific decisions that AI output would now feed, and it pushed those decisions to the level where the work and the context actually lived. The senior leaders kept the decisions that genuinely required them and gave up the ones they had been holding out of habit. Throughput of decisions rose to meet throughput of options.

Authority is the condition that converts generated possibility into organizational action. Without it, more capability produces a larger queue and a quieter workforce.

Adaptability: the debt either gets retired or it compounds

Everything AI produces has a maintenance cost attached. Code has to be understood by whoever inherits it. Processes have to be revised when they stop fitting. Documents have to be found, or they become clutter that costs attention every time someone searches past them.

Adaptability is whether the organization can retire what it built. It is not a synonym for moving fast. It is the capacity to unwind a decision, deprecate a process, and absorb the fact that something which worked last quarter no longer works.

James March framed the underlying tension in 1991 as the trade between exploration and exploitation. Organizations must both discover new possibilities and refine existing certainties, and the two compete for the same finite resources. AI dramatically lowers the cost of exploration. It does nothing at all to the cost of exploitation, which is where retirement, integration and maintenance live. An organization that only got cheaper at one half of that trade has not become more adaptable. It has become more lopsided, and the imbalance shows up as accumulated debt.

In the first organization, nothing was ever retired. Every AI initiative was additive. New tools on top of old tools, new reviews on top of old reviews, new dashboards next to the ones nobody had deprecated. Twelve months in, the organization was carrying everything it carried before plus everything it had added, and the marginal cost of each new thing was rising because each new thing had more to integrate with.

In the second organization, retirement was part of the work rather than an aspiration. Adopting a capability meant naming what it replaced, and the replacement was tracked with the same seriousness as the adoption. That is why the second organization could name what it stopped doing. It had been keeping the list on purpose.

The soil argument

The reason the same seed produces two harvests is that it went into two different soils, and nobody buying seed thinks to ask about the ground.

Chapter 4 of Builders Build makes this the organizing idea of the book: organizations produce what their conditions allow, and the soil determines what can grow. It is not a metaphor about culture and it is not a claim about effort. It is a claim about capacity. A seed carries potential, and the ground decides how much of that potential is expressible. Excellent seed in depleted ground produces a poor harvest, and the failure looks like the seed to anyone who was not looking at the dirt.

This is why the technology comparison keeps coming out inconclusive. The organizations getting compounding value and the organizations getting expensive activity are running the same models with the same context windows on the same infrastructure. The technical variables are close to identical. The variance lives entirely in the conditions.

There is precedent for the lag, and it argues for patience rather than panic. Robert Solow observed in 1987 that the computer age was visible everywhere except in the productivity statistics. The gap was real and it lasted years. It closed when organizations reorganized the work, not when better machines arrived. Brynjolfsson, Rock and Syverson revisited exactly this question for AI in 2017 and argued that the delay between a general purpose technology and its measured productivity effect is normal, and that it is caused by the time required to build the complementary organizational assets that let the technology pay off.

Those complementary assets have names. They are attention, alignment, authority and adaptability. Economists call them intangible capital. The book calls them soil. They are the same thing, and they behave the way soil behaves: slow to build, easy to deplete, and decisive.

What to fix first, and why the sequence matters

The four conditions are not a menu. They are ordered, and the order is not arbitrary.

Start with attention, because it is the only one that is self-blinding. An organization measuring consumption cannot see its alignment gaps, cannot see its decision backlog, and cannot see its accumulating debt, because none of those things appear on a usage dashboard. Fixing attention does not solve the other three. It makes them visible, and nothing else can be fixed while it is invisible. This is the cheapest of the four to change and the one most often skipped, because changing it means reporting a worse-looking number for a while.

Take alignment second, because it determines what the other repairs are aiming at. Pushing decisions downward is dangerous when the people receiving them are working from five different definitions of better. Alignment before authority, or the decentralization amplifies the incoherence.

Take authority third, once the definition is shared. This is where the visible gains arrive, because it is the condition that converts everything already being generated into things that actually happen. It is also the one that requires senior leaders to give something up, which is why it is usually attempted last and abandoned first.

Take adaptability fourth and treat it as permanent. It is not a project with an end date. It is the ongoing practice of retiring what no longer earns its keep, and it is the condition that determines whether the first three repairs hold or slowly erode.

None of this requires additional spend on tooling. Every organization I work with that is getting compounding value from AI is running roughly the same technology as the organizations that are not. What differs is the ground it was planted in, and the ground is something an organization can change without a purchase order.

AI does not create organizational problems. It inherits them. Scattered attention produces AI output that pulls in a dozen directions. Weak alignment produces results that are locally optimized and organizationally incoherent. Concentrated authority turns the technology into one more instrument of centralization. Brittle adaptability leaves an organization unable to learn at the speed its environment now demands. All of that was true before the tools arrived. The tools made it visible, expensive, and fast.

Which is the point the entire series has been circling.

The Four A’s are not a response to AI. They are the organizational readiness that AI requires.

From Builders Build, Chapter 9

Sources